Probabilistic forecasting and Bayesian modelling for planning and R&D decisions.
I’m Ellen Marsh, an engineer and statistician. I build forecasts and hierarchical Bayesian models for teams who have to act on them, and I deliver each one with the backtests, calibration checks and handover documents that show how far it can be trusted.
| Months ahead | 1 | 2 | 3 | 4 | 5 | 6 |
|---|---|---|---|---|---|---|
| Calibrated forecast, outturns inside its 90% range | 87.5% | 88.3% | 88.7% | 89.6% | 89.5% | 90.3% |
| The same forecast, uncalibrated | 74.9% | 72.0% | 71.9% | 70.8% | 69.3% | 68.3% |
nhs-ae-forecast/results/H-confirmatory/tables/primary.table.csv.
A forecast that says 90% should be right about nine times in ten. Uncalibrated, this one was right about seven. With a calibration layer built from its own recent errors, it held the registered tolerance at every horizon on a window it had never touched, and cut winter error by 38–48% against the seasonal-naive baseline. Read the case study.
Next public forecast: A&E attendances and emergency admissions for every integrated care board in England, December 2026 to March 2027, published by 31 October 2026 and scored against each month’s outturn as it is released, from 12 November.
Projects
Three pieces of work. For each I registered the claim before looking at results, reported the baselines before the candidates, measured the coverage of the intervals rather than assuming it, and expressed the answer in the units the decision uses. Each repository installs and tests in two commands. Where a simpler model won, the page names it and gives the margin.
-
Attendances and emergency admissions for England, one to six months ahead, from every published version of the data since 2015. Pre-registered, sealed, and published live from 31 October 2026 for all 36 integrated care boards.87.5–90.3% of outturns inside the 90% range at every horizon; 68–75% without calibration
-
Three published cell-culture media campaigns replayed on their own measurements: a Bayesian surrogate refit every five experiments against random selection, plus a planning table for how to seed and batch the next campaign. The tool takes one CSV and returns a one-page report.45–65 vs 95–105 experiments to a top-1% medium, model-guided against random, on the campaigns’ own enriched pools
-
Eight published cell-culture media studies catalogued to one schema, each with a dataset card recording units, design, replicate structure, licence and the loading decisions applied, and a one-command loader.8 studies on one schema with a licence register; six mirrored, two as fetch instructions where the licence forbids it
Offers
I take on four kinds of engagement, each scoped in writing before it starts. Scope, duration and deliverables are on the offers page.
Forecast audit
You already produce a forecast and want to know whether to plan on it. I test it against a naive baseline on rolling origins, measure whether its ranges hold, look for leakage and for the effect of data revisions, and hand back a memo and a monitoring plan.
Probabilistic demand forecasting
For a planning decision that repeats and currently rests on a point forecast or last year’s number. Baselines first, then quantile forecasts on rolling-origin backtests, a decision table in your units, a scheduled pipeline, a handover pack and a one-page memo. The NHS project is this engagement run in public.
Campaign replay audit
Send the CSV of a screening campaign. I replay it on its own measurements against random selection and tell you how many experiments a model-guided loop would have saved, and how to seed and batch the next one.
Bayesian measurement
For questions where a single number would mislead and the data arrives from many units at once: what drives the response and where it saturates, how demand responds to price across stores, what a new unit can borrow from the ones before it. Hierarchical models in PyMC, with an identifiability audit, predictive checks and calibration against the experiments you already have.
Two days a week from start month TBD, remote from the UK. £500 a day, or a fixed fee against a written scope. Inside or outside IR35.
I also take one-hour expert calls and run half-day workshops on forecasting practice, Bayesian workflow and bioprocess modelling.
Send a scope or a question, or email [email address TBD].